Add an accepted column so a usable fit reads as a predicate

[matthewholman]

Closes #498.

Reviewers on #495 asked for a status column that reads as a predicate -- if (!error) or if (success) -- rather than an integer compared against a documented list.

The two are not interchangeable: 0 means converged today, so !error works with the values unchanged while success is false for every good fit unless the values also invert. And renaming flag is breaking -- seven fixture CSVs carry it as a column header, so output already written stops loading.

This is the alternative from #498 instead: keep flag, add a derived boolean. accepted is 1 exactly when flag == 0. It reads the same, inverts nothing, and does not label -1 (never attempted) or 8 (incremental bookkeeping) as errors, which they are not.

It is derived from the flag the row carries rather than tracked beside the other outcome facts, so the two cannot disagree, and it is taken after any stage marker has overwritten the fitter's own verdict -- that overwritten value is what a reader sees.

Two things came along with it:

• create_empty_result hand-copied the outcome tuple, a second copy of the column layout that would have silently desynced. It now calls as_row.
• Corrects the failed_cov entry in docs/fit_flags.rst. Both of its assignments sit inside the non-gravitational weak-constraint guard, so a gravity-only fit is never marked there. The page implied the covariance was examined for every fit, which would let a reader take a column of zeros as reassurance.